Bayesian Networks and
Decision-Theoretic Reasoning
for Artificial Intelligence

September 1997

Click here to start

Table of Contents

Bayesian Networks and
Decision-Theoretic Reasoning
for Artificial Intelligence

Overview

Science- AAAI-97

Applications

Teenage Bayes

Course Contents

Probabilities

Discrete Random Variables

Continuous Random Variable

More Probabilities

Rules of Probability

Bayes Rule

Course Contents

Bayesian networks

Bayesian Networks

Product Rule

Marginalization

Bayes Rule Revisited

A Bayesian Network

Independence

Conditional Independence

More Conditional Independence:
Naïve Bayes

Naïve Bayes in general

More Conditional Independence:
Explaining Away

Put it all together

General Product (Chain) Rule
for Bayesian Networks

Conditional Independence

Another non-descendant

Independence and Graph Separation

Bayesian networks

Nodes as functions

PPT Slide

Causal Independence

Fine-grained model

Noisy-Or model

PPT Slide

Context-specific Dependencies

Asymmetric dependencies

Asymmetric Assessment

Continuous variables

Gaussian (normal) distributions

Gaussian networks

Composing functions

PPT Slide

Bayesian Networks

What is a variable?

Clarity Test:
Knowable in Principle

Structuring

Do the numbers really matter?

Bayesian Networks and Structure

Course Contents

Inference

Predictive Inference

Combined

Explaining away

Inference in Belief Networks

Basic Inference

Product Rule

Marginalization

Basic Inference

Inference in trees

Polytrees

The problem with loops

The problem with loops contd.

Variable elimination

Inference as variable elimination

Variable Elimination with loops

Join trees*

Exploiting Structure

Noisy-or decomposition

Inference with continuous variables

Computational complexity

Stochastic simulation

Likelihood weighting

Other approaches

PPT Slide

Course Contents

Decision making

Decision making

A Decision Problem

Value Function

Preference for Lotteries

Desired Properties for Preferences over Lotteries

Expected Utility

Some properties of U

Attitudes towards risk

Are people rational?

Maximizing Expected Utility

Multi-attribute utilities
(or: Money isn’t everything)

Influence Diagrams

Decision Making with Influence Diagrams

Value-of-Information

Value-of-Information in an
Influence Diagram

Value-of-Information is the increase in Expected Utility

Course Contents

Learning networks from data

The learning task

Parameter learning: one variable

Maximum likelihood

Bayesian approach

Conditioning on data

PPT Slide

General parameter learning

Partially observable data

Intuition

Expectation Maximization (EM)

Structure learning

Search space

Heuristic search

Scoring

Better scoring functions

Hidden variables

Randomly scattered data

Actual data

Bayesian clustering (Autoclass)

Clustered distributions

Detecting hidden variables

Course Contents

Reasoning over time

Dynamic environments

Dynamic Bayesian networks

Hidden Markov model

Hidden Markov models (HMMs)

HMMs and DBNs

Acting under uncertainty

Partially observable MDPs

Structured representation

Causality

Causal Theory

Setting vs. Observing

Predicting the effects of
interventions

Mechanism Nodes

Persistence

Course Contents

Applications

Why use Bayesian Networks?

Pathfinder

Studies of Pathfinder Diagnostic Performance

Commercial system: Integration

On Parenting: Selecting problem

On Parenting : MSN

Single Fault approximation

On Parenting: Selecting problem

Performing diagnosis/indexing

RICOH Fixit

FIXIT: Ricoh copy machine

Online Troubleshooters

Define Problem

Gather Information

Get Recommendations


Vista Project: NASA Mission Control

Costs & Benefits of Viewing Information

Status Quo at Mission Control

Time-Critical Decision Making

Simplification: Highlighting Decisions

Simplification: Highlighting Decisions

Simplification: Highlighting Decisions

What is Collaborative Filtering?

Bayesian Clustering for Collaborative Filtering

Applying Bayesian clustering

MSNBC Story clusters

Top 5 shows by user class

Richer model

What’s old?

What’s new?

Some Important AI Contributions

What’s in our future?

Authors: Jack Breese, Daphne Koller

Email: breese@microsoft.com koller@robotics.stanford.edu